Zero-Shot Context Generalization in Reinforcement Learning from Few Training Contexts